A short-term power load forecasting method considering load volatility and double-frequency modeling

Abstract

The application relates to the technical field of power system load prediction, and discloses a short-term power load prediction method considering load fluctuation and double-frequency modeling, which comprises the following steps: acquiring power load historical data of a target area in a predetermined historical period and contemporaneous influence factor data, performing time scale alignment and screening on the data to obtain a processed data set, decomposing the load sequence in the processed data set, performing feature enhancement on input data through a feature extraction model, and outputting an enhanced feature sequence; accurately anchoring the overall trend of the load through a linear regression model, and fully mining the dependency relationship of the residual sequence through a time series prediction model; and finally obtaining a final load prediction result through double-channel output superposition reconstruction. The application has the advantages that meteorological factors, dates and other multi-dimensional factors are comprehensively considered, and the STL load decomposition technology and the double-frequency model are used to significantly improve the precision and robustness of short-term power load prediction.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and in particular to a short-term power load forecasting method that takes into account load fluctuations and dual-frequency modeling. Background Technology

[0002] Power load forecasting has important theoretical value and practical significance in the planning and operation of new power systems. As a core component of the energy management system, high-precision forecasting is not only crucial for the safe and stable operation of the power system, but also an important foundation for the optimal allocation of resources and economic dispatch. Reference [1] summarizes the classic power load forecasting methods, such as time series method, regression analysis method, exponential smoothing method, and Kalman filter method. These methods consider the temporal nature of the data, but they have limitations in the processing and forecasting of nonlinear relational data. Reference [2] uses continuous and discrete time series combined with micro and macro information to generate multiple time series, and then uses an RNN model to learn the features in the multiple time series, thereby improving the performance of short-term power load forecasting. Reference [3] performs one-hot encoding of weekday and holiday information, constructs a 4-dimensional tensor in combination with load data, and then uses LSTM for load forecasting. The performance of the proposed forecasting model is verified on a public dataset. Reference [4] proposed a load forecasting model with a 5-layer network structure, including an input layer, two GRU layers, a normalization layer and an output layer. The dropout strategy was adopted in the training process of the network. The experimental results verified the performance of the model in various forecasting scenarios. Reference [5] proposed an LSTM load forecasting model based on an ensemble strategy. First, multiple training subsets were constructed using the Bootstrap sampling method. Load forecasting models were constructed on each subset. Finally, the results of all forecasting models were weighted and summed to obtain the final forecasting result. Reference [6] used a CNN model to extract features from the input data and capture seasonal cycles for load forecasting. The accuracy was higher on highly nonlinear sequences such as load data. The above research results have certain advantages in the field of load forecasting, but there are the following shortcomings. First, most existing studies only focus on the characteristics of power load and lack attention to meteorological influencing factors, and cannot comprehensively examine the important features in load forecasting. Second, existing studies focus on the optimization and improvement of deep learning models and ignore the importance of feature extraction at the front end of the model, resulting in a certain loss of forecasting accuracy.

[0003] References

[0004] [1] Ahmad T, Chen H, Guo Y, et al. A comprehensive overview on thedata driven and large scale based approaches for forecasting of buildingenergy demand: A review[J]. Energy and Buildings. 2018, 165: 301-320.

[0005] [2] Zhang B, Wu J, Chang P. A multiple time series-based recurrentneural network for short-term load forecasting[J]. Soft Computing. 2018, 22:4099-4112.

[0006] [3] Kong W, Dong Z Y, Jia Y, et al. Short-term residential loadforecasting based on LSTM recurrent neural network[J]. IEEE transactions onsmart grid. 2017, 10(1): 841-851.

[0007] [4] Wen L, Zhou K, Yang S. Load demand forecasting of residentialbuildings using a deep learning model[J]. Electric Power Systems Research.2020, 179: 106073.

[0008] [5] Tan M, Yuan S, Li S, et al. Ultra-short-term industrial powerdemand forecasting using LSTM based hybrid ensemble learning[J]. IEEEtransactions on power systems. 2019, 35(4): 2937-2948.

[0009] [6] Hu Z, Bao Y, Xiong T. Comprehensive learning particle swarmoptimization based memetic algorithm for model selection in short-term loadforecasting using support vector regression[J]. Applied Soft Computing. 2014,25: 15-25. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a short-term power load forecasting method that takes into account load fluctuation and dual-frequency modeling.

[0011] The objective of this invention is achieved through the following technical solution: a short-term power load forecasting method considering load fluctuation and dual-frequency modeling, the method comprising,

[0012] S1. Data processing: Obtain historical power load data of the target area within a predetermined historical period, as well as data on influencing factors during the same period. Perform time-scale alignment and filtering on the data to obtain the processed dataset. Through the nested inner and outer loop iteration process of STL, decompose the load sequence in the processed dataset to obtain trend components, seasonal components, and residual components.

[0013] S2. Feature extraction: Input the residual components after load decomposition into the CNN feature extraction model, perform feature enhancement on the input data through the feature extraction model, and output the enhanced feature sequence.

[0014] S3. Dual-frequency prediction: Construct a dual-channel parallel prediction architecture, which includes a linear channel and a deep learning channel. The linear channel accurately anchors the overall trend of the load by inputting the trend component and seasonal component into the linear regression model. The deep learning channel utilizes the enhanced feature sequence output by the CNN feature extraction model to fully explore the dependency relationship of the residual sequence through the time series prediction model. The final load prediction result is obtained by superimposing and reconstructing the outputs of the two channels.

[0015] Specifically, the influencing factors include date and meteorological factors. The date factor includes weekday and holiday indicators, and the meteorological factors include ambient temperature, humidity, wind speed, and perceived temperature. Based on the acquired temperature, humidity, and wind speed data, the perceived temperature is calculated as follows:

[0016] ;

[0017] ;

[0018] In the formula, The perceived temperature; For temperature; It is the water vapor pressure; The wind speed is at a height of 10 meters above the ground. This refers to relative humidity.

[0019] Specifically, the screening and preprocessing of influencing factor data includes analyzing the impact of meteorological factors on the load using the Pearson correlation coefficient, and identifying influencing factors highly correlated with the load:

[0020] ;

[0021] In the formula, The linear correlation coefficient between the input variables; and There are two variables; and for and Their respective averages;

[0022] The selected influencing factor data, historical electricity load data, and periodic time feature vectors generated based on date factors are concatenated to generate the processed dataset.

[0023] ;

[0024] ;

[0025] In the formula, for Historical load at any moment; Input for each dimension; The total dimension of the features; for The time is a weekday; for Every moment is a holiday.

[0026] Specifically, decomposing the load sequences in the processed dataset includes:

[0027] The decomposition model is defined as follows:

[0028] ;

[0029] In the formula: For a moment The original power load observations; For trend components; For seasonal quantities; These are the residual components;

[0030] The LOESS smoother is used to locally fit the trend and seasonal terms. For each target point, LOESS fits a low-order polynomial in its local neighborhood by minimizing the weighted least squares error.

[0031] ;

[0032] In the formula: This is the target time point where smoothing estimation is needed. For The time position of sample points within a local window centered on the sample point; This represents the total number of samples contained within the local window; It is a low-order polynomial fitted within the local neighborhood; For weighting functions; For bandwidth parameters; For time points The actual load value.

[0033] Specifically, the feature extraction model includes a one-dimensional convolutional layer and a function activation layer. It performs convolution operations on the input sequence using a set of learnable convolutional kernels. Each convolutional kernel slides along the time axis, outputting an enhanced feature sequence:

[0034] ;

[0035] ;

[0036] In the formula, The linear output of the convolutional layer; For the first One convolutional kernel; This refers to the temporal position of the convolution output; For the first Bias terms of each convolution kernel; For feature dimensions; For the first Each convolutional kernel in the channel Offset within window Weight at each location; For the first Variables in time The value of ; is the kernel length.

[0037] Specifically, in S3, a linear regression model is used to directly map the future trend to the sum of the seasons:

[0038]

[0039] In the formula, These are the predicted values ​​for the linear components; It is a linear weight matrix; Weather characteristics for a future moment; This is a bias term.

[0040] Specifically, the time-series prediction model includes an LSTM layer, a Dropout layer, and a fully connected layer. The LSTM layer processes and enhances the feature sequence, and through its internal gating mechanism, captures long-term dependencies throughout the sequence, outputting a final state vector. The principle formula of the LSTM layer is expressed as follows:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] In the formula, , , , , and These are the forget gate, input gate, candidate memory cell state vector, memory cell state vector, output gate, and hidden state vector at time step t, respectively. Let be the input vector at time step t; In order to take steps in the previous time The hidden state vector; In order to take steps in the previous time The state vector of the memory cell; These are the weight matrices corresponding to the forget gate, input gate, memory unit, and output gate, respectively. These are the bias vectors corresponding to the forget gate, input gate, memory unit, and output gate, respectively. To hide the previous time step's state vector and the input vector at the current time step The new vector formed by splicing; sigmoid is the activation function; tanh is the hyperbolic tangent activation function; The Hadamard product is used; the state vector is input into the fully connected layer, mapping the high-dimensional feature vector to the predicted values ​​of the residual components. ;

[0049] The final load forecast result is obtained by superimposing and reconstructing the dual-channel outputs:

[0050] ;

[0051] In the formula, This is the final predicted value; These are the predicted values ​​for the linear components; These are the predicted values ​​for the residual components.

[0052] The present invention has the following advantages:

[0053] In the data processing stage, this invention utilizes physical property analysis and Pearson correlation analysis to screen meteorological and date factors, constructing an input dataset, and employs STL decomposition to perform trend decomposition of the load. In the feature extraction and enhancement stage, the dataset is processed using a time sliding window approach to generate continuous feature maps containing spatiotemporal information. These feature maps are then fed into a one-dimensional convolutional neural network to effectively extract deep-seated potential correlations between consecutive time points and different feature dimensions within the feature maps, forming a series of feature vectors with higher information density. In the prediction stage, a linear regression model predicts low-frequency trends. For the residual components, a pre-processed CNN model enhances and refines the features of the original multi-dimensional time-series data, providing a large amount of structured, high-quality, and effective input for the subsequent LSTM network model, significantly improving the model's learning efficiency and prediction accuracy. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the load forecasting method of the present invention;

[0055] Figure 2 A heatmap showing the correlation of factors affecting electricity load. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0057] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0058] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0059] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0060] like Figures 1 to 2 As shown, a short-term power load forecasting method considering load volatility and dual-frequency modeling is proposed. This method includes:

[0061] S1. Data processing: Obtain historical power load data of the target area within a predetermined historical period, as well as data on influencing factors during the same period. Perform time-scale alignment and filtering on the data to obtain the processed dataset. Through the nested inner and outer loop iteration process of STL, decompose the load sequence in the processed dataset to obtain trend components, seasonal components, and residual components.

[0062] The factors influencing electricity load are multifaceted. When considering these factors, a comprehensive analysis is essential to ensure no important element is overlooked. Short-term load forecasting factors can be categorized into date and meteorological factors. Date factors include weekday and holiday markers. Date information is a crucial factor influencing the inherent periodicity of electricity load, with load changes exhibiting strong periodicity across different time scales. Within a seven-day week, electricity load displays a clear periodicity, and load patterns also emerge during statutory holidays. Statutory holidays lead to significant changes in social production and lifestyle patterns, resulting in load characteristics that differ from ordinary weekdays or weekends. Meteorological factors have a significant impact on electricity load. Common meteorological factors affecting load include temperature, humidity, wind speed, and weather type. Compared to ambient temperature in weather forecasts, perceived temperature more accurately reflects residents' electricity consumption behavior. Based on acquired temperature, humidity, and wind speed data, perceived temperature is calculated as follows:

[0063] ;

[0064] ;

[0065] In the formula, The perceived temperature; For temperature; It is the water vapor pressure; The wind speed is at a height of 10 meters above the ground. This refers to relative humidity.

[0066] The screening and preprocessing of influencing factor data includes,

[0067] Factors with a higher correlation to electricity load than a preset threshold were identified based on the Pearson correlation coefficient.

[0068] ;

[0069] In the formula, The linear correlation coefficient between the input variables; and There are two variables; and for and Their respective averages;

[0070] The selected influencing factor data, historical electricity load data, and periodic time feature vectors generated based on date factors are concatenated to generate the processed dataset.

[0071] ;

[0072] ;

[0073] In the formula, for Historical load at any moment; Input for each dimension; The total dimension of the features; for The time is for weekdays. for Every moment is a holiday.

[0074] Based on LOESS seasonal trend decomposition, electricity load data is typically composed of fluctuations at multiple different frequencies. STL decomposition, as a non-parametric filtering process, decomposes the load sequence in the processed dataset through nested inner and outer loop iterations.

[0075] ;

[0076] In the formula: For a moment The original power load observations; The trend component reflects the long-term changing trend of low-frequency components in the load sequence; The seasonal component reflects a fixed fluctuation pattern with a period of a specific time length; The residual component reflects the remaining portion after removing trends and seasonality.

[0077] The core mathematical mechanism of STL lies in using the LOESS smoother to locally fit the trend and seasonal terms. For each target point, LOESS fits a low-order polynomial in its local neighborhood by minimizing the weighted least squares error, as shown in the following formula:

[0078] ;

[0079] In the formula: This is the target time point where smoothing estimation is needed. For The time position of sample points within a local window centered on the sample point; This represents the total number of samples contained within the local window; It is a low-order polynomial fitted within the local neighborhood; The weighting function assigns weights based on the distance between sample points and target points; For bandwidth parameters; For time points The actual load value.

[0080] S2. Feature extraction: Input the residual components after load decomposition into the CNN feature extraction model, perform feature enhancement on the input data through the feature extraction model, and output the enhanced feature sequence.

[0081] The feature extraction model comprises two stacked one-dimensional convolutional layers and one function activation layer. The first convolutional layer has a kernel size of 3, 32 filters, Same padding, and ReLU activation. This layer is used to capture fine-grained fluctuation patterns. The second convolutional layer maintains the kernel size of 3 but increases the number of filters to 64 to extract higher-dimensional abstract features. A batch normalization layer is introduced to accelerate convergence. The CNN convolutional neural network utilizes the sliding window mechanism of the convolutional kernel to effectively capture local features of the data. For multivariate time-series data such as electricity load, one-dimensional convolution is used to extract local dependency patterns in the time dimension. A set of learnable convolutional kernels is used to perform convolution operations on the input sequence. The nth convolutional kernel slides along the time axis, and the nth output feature map of it... The formula for calculating each element is shown in the figure, and can be configured... With different convolutional kernels, the network can extract... For different feature patterns, the calculation of the output enhanced feature sequence is as follows:

[0082] ;

[0083] ;

[0084] In the formula, The linear output of the convolutional layer; For the first One convolutional kernel; This refers to the temporal position of the convolution output; For the first Bias terms of each convolution kernel; For feature dimensions; For the first Each convolutional kernel in the channel Offset within window Weight at each location; For the first Variables in time The value of ; is the kernel length.

[0085] This invention integrates multiple heterogeneous time series affecting workload—including historical workload, five-dimensional feature data, and two constructed periodic time features. Borrowing from word vector representation in natural language processing, all relevant features at each time point are concatenated into a single feature vector. The original multiple independent time series are coupled into a novel, unified multidimensional time series. Subsequently, a sliding window approach is used to extract continuous segments from this new multidimensional time series to construct the model's input samples. In this invention, the width of the sliding window is set to 24 time steps, and the step size is 24. Therefore, each input sample is a two-dimensional feature map of size 8×24, where 8 represents the number of features and 24 represents the time step length. A one-dimensional convolutional layer of size 5 containing 64 kernels is used. After processing, the original 8×24 feature map is transformed into a higher-level, 64×24 enhanced feature sequence. Each time step of this new sequence incorporates contextual information from its neighboring time points, resulting in a higher information density than the original input.

[0086] S3. Dual-frequency prediction: Construct a dual-channel parallel prediction architecture, which includes a linear channel and a deep learning channel. The linear channel accurately anchors the overall trend of the load by inputting the trend component and seasonal component into the linear regression model. The deep learning channel utilizes the enhanced feature sequence output by the CNN feature extraction model to fully explore the dependency relationship of the residual sequence through the time series prediction model. The final load prediction result is obtained by superimposing and reconstructing the outputs of the two channels.

[0087] In the linear regression model, the trend and seasonal components encompass low-frequency long-term trends and fixed-period patterns in the load data. These signals are typically stable and highly regular, and overly complex deep networks can easily lead to overfitting. Therefore, a linear regression channel is constructed, using future meteorological forecasts as explanatory variables, and a multi-output linear regression model is used to directly map the sum of future trends and seasons.

[0088] ;

[0089] In the formula, These are the predicted values ​​for the linear components; It is a linear weight matrix; Weather characteristics for future moments, This is a bias term.

[0090] The residual component contains high-frequency nonlinear fluctuations caused by sudden events or random electricity consumption, which is a key bottleneck for prediction accuracy. The time-series prediction model includes LSTM layers, Dropout layers, and fully connected layers. The LSTM layer processes and enhances the feature sequence, and through its internal gating mechanism, captures long-term dependencies throughout the sequence, outputting a final state vector. The forgetting gate is crucial for LSTM to achieve long-term memory. It determines the state based on the output of the previous time step. and the input at the current moment The sigmoid function determines which memory cell should be used. The sigmoid function determines which historical information is forgotten or retained. Its output value ranges between 0 and 1, where 0 represents "complete forgetting" and 1 represents "complete retention." The input gate decides which new information will be stored in the memory unit at the current moment. It consists of three parts: first, a sigmoid layer determines which values ​​need to be updated; second, a tanh layer generates a vector of candidate update values. Finally, the two are multiplied to determine the new information to be written to the memory cell. Combining the forget gate and the input gate, LSTM updates the state of the memory cell. It first "forgets" some old information through the forget gate, then "remembers" some new information through the input gate, ultimately obtaining the new cell state at the current moment. The output gate ultimately determines the output at the current moment. It first uses a sigmoid layer to determine which parts of the memory cell will be output, and then processes the state of the memory cell. The data is preprocessed using a tanh function, and then multiplied together to obtain the final output. This is both the prediction result at the current moment and will be passed on as information to the next time step; specifically, it is expressed as:

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] In the formula, , , , , and These are the forget gate, input gate, candidate memory cell state vector, memory cell state vector, output gate, and hidden state vector at time step t, respectively. In time step The input vector; In order to take steps in the previous time The hidden state vector; In order to take steps in the previous time The state vector of the memory cell; These are the weight matrices corresponding to the forget gate, input gate, memory unit, and output gate, respectively. These represent the bias vectors corresponding to the forget gate, input gate, memory unit, and output gate, respectively. To hide the previous time step's state vector and the input vector at the current time step The new vector formed by splicing; sigmoid is the activation function; tanh is the hyperbolic tangent activation function; The Hadamard product is used; the state vector is input into the fully connected layer, mapping the high-dimensional feature vector to the predicted values ​​of the residual components. ;

[0099] The final load forecast result is obtained by superimposing and reconstructing the dual-channel outputs:

[0100] ;

[0101] In the formula, This is the final predicted value; These are the predicted values ​​for the linear components; These are the predicted values ​​for the residual components.

[0102] When training the model, the original dataset is divided into a training set, a validation set, and a test set. The training set is used to train the model, the validation set is used to verify the training effect, and the test set is used to measure the model's prediction performance. The division method is as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] In the formula, , , These are the training set, validation set, and test set, respectively. For the first One input sample; In order to be with the first The target output value corresponding to each input sample, that is, the actual power load value at the time to be predicted; , , These represent the total number of samples in the training set, validation set, and original dataset, respectively.

[0107] To eliminate the negative impact of dimensional differences between different features and improve the training efficiency and performance of the model, this study normalized all input features and output targets before feeding the data into the model. In this study, the min-max normalization method was used. This method maps all values ​​to the interval [0,1] by performing a linear transformation on the original data. Its mathematical formula is as follows:

[0108] ;

[0109] In the formula, These are the original data points; and These are the minimum and maximum values ​​in the feature sequence, respectively. That is the normalized value.

[0110] After the model completes its prediction, the same set of parameters obtained from the training set is used to inversely normalize the model's output.

[0111] ;

[0112] To effectively evaluate the prediction performance, the evaluation metrics are introduced: root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] In the formula, For the first The actual load value at each time point; This represents the model's predicted load value for that time point. For the first Prediction error at each time point This represents the total number of samples in the test set.

[0118] Mean squared error was used as the loss function, the Adam algorithm was used for parameter optimization, and the cross-correlation function and peak lag function were used to evaluate the prediction delay effect.

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] In the formula, The loss function; The regularization coefficient is used. The learning rate; and These are the bias correction values ​​for the first-order moment estimate and the second-order moment estimate of the gradient, respectively; To prevent smooth terms with a denominator of zero; The displacement is the lag displacement (unit: hours). and These represent the mean and standard deviation of the sequence, respectively.

[0125] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.

Claims

1. A short-term power load forecasting method considering load fluctuation and dual-frequency modeling, characterized in that: The method includes, S1. Data processing: Obtain historical power load data of the target area within a predetermined historical period, as well as data on influencing factors during the same period. Perform time-scale alignment and filtering on the data to obtain the processed dataset. Through the nested inner and outer loop iteration process of STL, decompose the load sequence in the processed dataset to obtain trend components, seasonal components, and residual components. S2. Feature extraction: Input the residual components after load decomposition into the CNN feature extraction model, perform feature enhancement on the input data through the feature extraction model, and output the enhanced feature sequence. S3. Dual-frequency prediction: A dual-channel parallel prediction architecture is constructed, which includes a linear channel and a deep learning channel. The linear channel accurately anchors the overall trend of the load by inputting the trend component and seasonal component into the linear regression model. The deep learning channel utilizes the enhanced feature sequence output by the CNN feature extraction model and fully explores the dependency relationship of the residual sequence through the time series prediction model. The final load prediction result is obtained by superimposing and reconstructing the outputs of the two channels. In S3, a linear regression model is used to directly map the future trend to the sum of the seasons: ； In the formula, These are the predicted values ​​for the linear components; It is a linear weight matrix; Weather characteristics for a future moment; This is a bias term.

2. The short-term power load forecasting method considering load fluctuation and dual-frequency modeling according to claim 1, characterized in that: The influencing factors include date and meteorological factors. The date factor includes weekday and holiday indicators, and the meteorological factors include ambient temperature, humidity, wind speed, and perceived temperature. Based on the acquired temperature, humidity, and wind speed data, the perceived temperature is calculated as follows: ； ； In the formula, The perceived temperature; For temperature; It is the water vapor pressure; The wind speed is at a height of 10 meters above the ground. This refers to relative humidity.

3. The short-term power load forecasting method considering load fluctuation and dual-frequency modeling according to claim 1, characterized in that: The screening and preprocessing of influencing factor data includes analyzing the impact of meteorological factors on the load using the Pearson correlation coefficient, and identifying influencing factors highly correlated with the load: ； In the formula, The linear correlation coefficient between the input variables; and There are two variables; and for and Their respective averages; The selected influencing factor data, historical electricity load data, and periodic time feature vectors generated based on date factors are concatenated to generate the processed dataset. ； ； In the formula, for Historical load at any moment; Input for each dimension; The total dimension of the features; for The time is a weekday; for Every moment is a holiday.

4. The short-term power load forecasting method considering load fluctuation and dual-frequency modeling according to claim 3, characterized in that: Decomposing the load sequences in the processed dataset includes: The decomposition model is defined as follows: ； In the formula: For a moment The original power load observations; For trend components; For seasonal quantities; These are the residual components; The LOESS smoother is used to locally fit the trend and seasonal terms. For each target point, LOESS fits a low-order polynomial in its local neighborhood by minimizing the weighted least squares error. ； In the formula: This is the target time point where smoothing estimation is needed. For The time position of sample points within a local window centered on the sample point; This represents the total number of samples contained within the local window; It is a low-order polynomial fitted within the local neighborhood; For weighting functions; For bandwidth parameters; For time points The actual load value.

5. A short-term power load forecasting method considering load fluctuation and dual-frequency modeling according to claim 4, characterized in that: The feature extraction model includes a one-dimensional convolutional layer and a function activation layer. It performs convolution operations on the input sequence using a set of learnable convolutional kernels. Each convolutional kernel slides along the time axis, outputting an enhanced feature sequence: ； ； In the formula, The linear output of the convolutional layer; For the first One convolutional kernel; This refers to the temporal position of the convolution output; For the first Bias terms of each convolution kernel; For feature dimensions; For the first Each convolutional kernel in the channel Offset within window Weight at each location; For the first Variables in time The value of ; is the kernel length.

6. The short-term power load forecasting method considering load fluctuation and dual-frequency modeling according to claim 1, characterized in that: The time-series prediction model includes an LSTM layer, a Dropout layer, and a fully connected layer. The LSTM layer processes and enhances the feature sequence, and through its internal gating mechanism, captures long-term dependencies throughout the sequence, outputting a final state vector. The principle formula of the LSTM layer is expressed as follows: ； ； ； ； ； ； ； In the formula, , , , , and At time step The forget gate, input gate, candidate memory cell state vector, memory cell state vector, output gate, and hidden state vector; In time step The input vector; In order to take steps in the previous time The hidden state vector; In order to take steps in the previous time The state vector of the memory cell; These are the weight matrices corresponding to the forget gate, input gate, memory unit, and output gate, respectively. These are the bias vectors corresponding to the forget gate, input gate, memory unit, and output gate, respectively. To hide the previous time step's state vector and the input vector at the current time step The new vector formed by splicing; sigmoid is the activation function; tanh is the hyperbolic tangent activation function; The Hadamard product is used; the state vector is input into the fully connected layer, mapping the high-dimensional feature vector to the predicted values ​​of the residual components. ; The final load forecast result is obtained by superimposing and reconstructing the outputs from both channels: ； In the formula, This is the final predicted value; These are the predicted values ​​for the linear components; These are the predicted values ​​for the residual components.

Images